What is AI referral traffic?

The expression groups visits of very different origin under one convenient label. A link tapped inside a conversation with an assistant, a link opened from that same assistant's mobile app, and a link copied from an answer and pasted into another browser all produce the same arrival at the server, yet they are recorded in three different ways.

The boundary with Search is worth drawing clearly. Clicks coming from Google's generative summary and conversational mode are classified as organic search rather than assistant traffic, because they originate in Search itself. When people talk about AI traffic they usually mean the conversational assistants, which are external sites and therefore arrive as a referral.

The term has no technical definition of its own either. No specification and no tool fixes it in any binding way; it describes a market observation that has settled through use. In practice this means two reports on the same site can produce different figures depending on which domains each one included and how each one treats visits without a referrer. Before comparing two figures, the domain list and the assignment rule behind each of them belong on the table.

Google Analytics 4 gained a dedicated channel for this in May 2026. When the referrer matches its list of recognised assistants, the session receives the medium ai-assistant and the campaign (ai-assistant), and appears in the AI Assistant channel of the default channel group. Three limits are worth keeping in mind: the full list of recognised domains is not published, the classification is not applied retroactively, and the channel expressly excludes the generative features of Search itself, which continue to count as organic search.

The rest arrives in worse shape. If the link opens in the in-app browser of a mobile application, or if the user copies the address and pastes it, the referrer header does not travel and the session lands in Direct with no trace of its origin. Some products append a utm_source parameter to the links they display, which rescues part of those cases, but that is product behaviour which can change and it does not cover addresses copied by hand.

To see it in your own account without depending on the automatic channel, create a custom channel group in the property administration, under the data settings. The useful condition combines two parts: session source against a regular expression holding the assistant domains you care about, and utm_source against that same list for the tagged links. After that it pays to review the acquisition report by referral every few weeks, because new domains appear regularly and the expression ages.

The decision that depends on this figure is whether a channel deserves investment and at what size. The risk runs in two directions. If you count only what arrives identified, the channel looks smaller than it is. If you attribute every rise in direct traffic to AI, it looks larger. Neither reading holds up, and both get used to justify budgets.

From this follows an uncomfortable warning for anyone selling work in this field. Plenty of share and growth percentages circulate about this channel whose collection is undocumented. A good part comes from panels that explain neither their size, nor their composition, nor what they did with visits lacking a referrer. A percentage with no method behind it is not data, and using it to size an investment moves the error into the budget.

What does support a decision is your own time series. A site can measure how many sessions identified as assistants it receives, which landing pages concentrate them, and how they behave against other channels on the conversions it has already defined. That figure is partial, and saying so is part of the job, but it is your own, comparable with itself over time, and reproducible by anyone who opens the same account.

• Create a custom channel group with a regular expression on session source that captures the assistant domains relevant to your site, and add a condition on utm_source to pick up the tagged links.
• Review the acquisition report by referral every month and add any new domains that show up to that expression, because this list ages faster than any other part of the configuration.
• Note the date you started measuring it and do not compare against periods before that date, since the automatic classification is not applied retroactively.
• Watch direct traffic alongside the assistant channel, and treat any relationship between the two as a hypothesis still to be confirmed rather than as an attribution.
• State in every report that the figure is an observed minimum and explain what falls outside it. A partial figure that is properly labelled is usable; the same figure presented as complete is not.
• Compare the behaviour of those sessions through landing pages and the conversions you already have defined, instead of stopping at volume, which is the part most affected by attribution gaps.

• Adding the assistant channel and the visits from Search's generative summary together as if they were the same thing. The latter are already counted in organic search and would be double counted.
• Reading every rise in direct traffic as unattributed AI. Direct also collects bookmarks, email clients, messaging apps and links pasted by hand.
• Using market share percentages with no published method to justify an investment, when it is unknown how many sites they come from or what was done with the sessions lacking a referrer.
• Comparing the channel across two different sites without checking that both use the same domain list and the same assignment rule.
• Building the entire report on the campaign parameter a product appends today, taking for granted that it will still be there tomorrow.